Neural network system, learning device, learning method and program

The neural network system with an integral-fire model and learning method to suppress neuron firing addresses high power consumption in spiking neural networks, achieving reduced power usage without compromising recognition accuracy.

JP2024167718APending Publication Date: 2024-12-04NEC CORP
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
JP2023083976
Authority / Receiving Office
JP · JP
Patent Type
Applications
Current Assignee / Owner
Filing Date
2023-05-22
Publication Date
2024-12-04

AI Technical Summary

Technical Problem

The high power consumption of spiking neural networks is a challenge that existing technologies have not adequately addressed.

Method used

A neural network system utilizing a time-based spiking neural network configured with an integral-fire model and a learning method that minimizes neuron firing probability through an evaluation function, reducing power consumption by training the network to suppress neuron model firing.

Benefits of technology

The system effectively reduces power consumption while maintaining recognition accuracy by training the network to minimize neuron model firing, achieving lower operational power usage.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 2024167718000001_ABST
    Figure 2024167718000001_ABST
Patent Text Reader

Abstract

To make power consumption relatively small when a spiking neural network is operating.SOLUTION: A neural network system comprises a time-based spiking neural network configured using a neuron model based on an integrate-and-fire model, and learning means for learning the spiking neural network using an evaluation function indicating a better evaluation as a probability that an individual neuron model fires decreases.SELECTED DRAWING: Figure 1
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] The present invention relates to a neural network system, a learning device, a learning method, and a program. [Background technology]

[0002] One type of neural network is a spiking neural network (SNN) (see, for example, Patent Document 1). [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Special Publication No. 2017-509953 Summary of the Invention [Problem to be solved by the invention]

[0004] It is preferable that a spiking neural network consumes as little power as possible when operating.

[0005] An example of an object of the present invention is to provide a neural network system, a learning device, a learning method, and a program that can solve the above-mentioned problems. [Means for solving the problem]

[0006] According to a first aspect of the present invention, a neural network system includes a time-based spiking neural network configured using neuron models based on an integrate-and-fire model, and a learning means for learning the spiking neural network using an evaluation function that indicates a better evaluation the lower the probability that each neuron model will fire.

[0007] According to a second aspect of the present invention, a learning device includes a learning means for learning a time-based spiking neural network configured using neuron models based on an integrate-and-fire model, using an evaluation function that indicates a better evaluation the lower the probability that each neuron model fires.

[0008] According to a third aspect of the present invention, a learning method includes a computer training a time-based spiking neural network configured using neuron models based on an integrate-and-fire model using an evaluation function that indicates a better evaluation the lower the probability that each neuron model in the network will fire.

[0009] According to a fourth aspect of the present invention, a program is provided for causing a computer to execute a process of learning a time-based spiking neural network configured using neuron models based on an integrate-and-fire model, using an evaluation function that indicates a better evaluation the lower the probability that each neuron model in the time-based spiking neural network will fire. Effect of the Invention

[0010] According to the present invention, it is expected that the power consumption during operation of a spiking neural network will be relatively small. [Brief description of the drawings]

[0011] [Figure 1] FIG. 1 is a diagram illustrating an example of a configuration of a neural network system according to an embodiment. [Diagram 2] FIG. 13 is a diagram showing an example of a time integral of a membrane potential in an embodiment. [Diagram 3] FIG. 11 is a diagram showing an example of a firing condition of a neuron model in the embodiment. [Figure 4] FIG. 13 is a diagram showing an example of the firing state of a neuron model observed when γ2=0 in an experiment related to an embodiment. [Diagram 5] FIG. 13 is a diagram showing an example of the firing state of a neuron model observed in an experiment related to an embodiment when γ2=2.3×10−7. [Figure 6] FIG. 13 is a diagram showing an example of the firing state of a neuron model observed in an experiment related to an embodiment when γ2=4.8×10−6. [Figure 7] FIG. 2 is a diagram showing a first example of the relationship between the recognition rate of a neural network and the firing rate of a neuron model in an intermediate layer in an embodiment. [Figure 8] FIG. 13 is a diagram showing a second example of the relationship between the recognition rate of a neural network and the firing rate of a neuron model in an intermediate layer in the embodiment. [Figure 9] FIG. 13 is a diagram showing a third example of the relationship between the recognition rate of a neural network and the firing rate of a neuron model in an intermediate layer in the embodiment. [Figure 10] FIG. 13 is a diagram showing a fourth example of the relationship between the recognition rate of a neural network and the firing rate of a neuron model in an intermediate layer in the embodiment. [Figure 11] FIG. 1 is a diagram illustrating an example of a configuration of a neural network system according to an embodiment. [Figure 12] FIG. 2 is a diagram illustrating an example of a configuration of a learning device according to an embodiment. [Figure 13] FIG. 11 is a diagram illustrating an example of a processing procedure in a learning method according to the embodiment. [Figure 14] FIG. 1 illustrates an example configuration of a computer in accordance with at least one embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0012] The following describes embodiments of the present invention, but the following embodiments do not limit the scope of the invention. Furthermore, not all of the combinations of features described in the embodiments are necessarily essential to the solution of the invention. In the following, a circumflexed letter may be represented by adding a ^ next to the letter. For example, a circumflexed v may also be represented as v^.

[0013] 1 is a diagram showing an example of the configuration of a neural network system according to an embodiment. In the configuration shown in FIG. 1, a neural network system 100 includes a spiking neural network 110 and a learning unit 120. The neural network system 100 is a system that performs learning of a spiking neural network 110. The neural network system 100 may be configured as one device, or may be configured as a combination of multiple devices. For example, the spiking neural network 110 may be configured using hardware including an analog circuit, and the learning unit 120 may be configured using a computer such as a personal computer (PC).

[0014] The spiking neural network 110 is a time-based spiking neural network configured using a neuron model (spiking neuron model) based on an integrate-and-fire (IF) model.

[0015] The spiking neural network referred to here is a neural network in which the neuron model outputs signals at timings based on a state quantity called membrane potential, which changes over time depending on the input status of signals to the neuron model itself. The signals input and output by the neuron model are also called spike signals, or simply spikes. The output of a spike by the neuron model is also called firing. The spiking neural network 110 uses a time method for transmitting information between neuron models, in which information is transmitted at the transmission timing of spike signals (firing times of neuron models).

[0016] In the integrate-and-fire model, the membrane potential changes in response to input spikes to the neuron model, and when the membrane potential reaches a threshold, the neuron model fires. The threshold used to determine the timing of firing of the neuron model is also called the threshold membrane potential or firing threshold.

[0017] The processing performed by spiking neural network 110 may be any processing that can be performed using a spiking neural network, including, but not limited to, image recognition, voice recognition, biometric authentication, or numerical prediction.

[0018] The learning unit 120 performs training on the spiking neural network 110. Specifically, the learning unit 120 updates parameter values ​​such as weighting coefficients for input spikes to each neuron model of the spiking neural network 110. In particular, the learning unit 120 performs training on the spiking neural network 110 using an evaluation function that indicates a better evaluation the lower the possibility that each neuron model of the spiking neural network 110 will fire. Furthermore, the learning unit 120 performs training on the spiking neural network 110 using a training method that uses the derivative of the evaluation function, such as backpropagation. The learning unit 120 corresponds to an example of a learning means.

[0019] Here, it is considered that power is consumed when the neuron model fires, such as when the neuron model outputs a voltage pulse as a spike. The power consumption of the spiking neural network 110 is expected to be smaller as the number of firings of the neuron model decreases (the firing frequency decreases). Therefore, by having the learning unit 120 train the spiking neural network 110 using an evaluation function that indicates a better evaluation the lower the possibility of the neuron model firing, it is expected that the number of times the neuron model fires will be relatively small, and the power consumption of the spiking neural network 110 will be relatively small.

[0020] In the following, an example will be described in which the learning unit 120 uses a loss function (cost function) as an evaluation function to learn the spiking neural network 110. The loss function here is a function in which the smaller the function value, the better the evaluation. However, the learning unit 120 may use a function in which the larger the function value, the better the evaluation. For example, the learning unit 120 performs learning of the spiking neural network 110 using a loss function C shown in equation (1).

[0021]

number

[0022] Σ p κ p lnS p is a term for improving the accuracy of estimation by the neural network, and equation (1) shows an example in which the neural network performs class classification. The term for improving the accuracy of estimation by the neural network is also called the estimation loss term. p indicates the identification number of a class. In the example of formula (1), it is assumed that the spiking neural network 110 has the same number of output nodes as the classes, and that the classes and the output nodes are in one-to-one correspondence. p is also used as the identification number of the output node.

[0023] κ p represents the teacher label. For the correct class, κ p = 1, and for other classes, κ p =0. ln represents the natural logarithm. S pis the softmax function used for the pth class, and is expressed as equation (2).

[0024]

number

[0025] exp represents the power of Napier's constant e. Like p, q indicates the class identification number. t p represents the firing time of the pth output node. σ soft is the softmax function S when the firing time of the output layer changes. p is a constant provided as a scale factor to adjust the degree of change in the value of σ soft >0. However, there is no particular limitation to the estimated loss term used by the learning unit 120. For example, the learning unit 120 may use a loss function in which a function other than a softmax function is used for the estimated loss term.

[0026] In equation (1), γ2V and γ3Q are examples of terms that indicate a better evaluation when the individual neuron models are less likely to fire. The sub-formula V is a sub-formula for evaluating the possibility of firing the neuron model based on the time integral of the membrane potential. The value of the sub-formula V is also called the time integral loss of the membrane potential. However, as described below, the value of the sub-formula V can be calculated without the need to obtain the value of the membrane potential. γ2 is a coefficient for adjusting the degree of influence of the value of the subexpression V on the value of the loss function C. The coefficient γ2 corresponds to an example of a coefficient for adjusting the degree of influence of a subexpression that shows a better evaluation when the probability that an individual neuron model fires is low. The value of the coefficient γ2 may be preset by, for example, a user.

[0027] The term γ2V, which is the time integral loss of the membrane potential, can be considered as a kind of regularization term. The regularization term here is a term that is set so that the value of the weight coefficient of the neural network becomes relatively small. The term γ2V, which is the time integral loss of the membrane potential, or the regularization by the term γ2V, which is the time integral loss of the membrane potential, is also called M-SSR (Membrane-potential-based Spike-timing-based sparse-firing regularization).

[0028] The subformula Q is a subformula for evaluating the possibility of the neuron model firing based on the firing condition. The value of the subformula Q is also called the loss for the firing condition. γ3 is a coefficient for adjusting the degree of influence of the value of the subexpression Q on the value of the loss function C. The coefficient γ3 corresponds to an example of a coefficient for adjusting the degree of influence of a subexpression that shows a better evaluation when the probability that an individual neuron model fires is low. The value of the coefficient γ3 may be preset by, for example, a user.

[0029] The loss term γ3Q related to the firing condition can be considered as a kind of regularization term. The loss term related to the firing condition γQ, or the regularization by the loss term related to the firing condition γQ, is also called F-SSR (Firing-condition-based SSR).

[0030] However, the loss function used by the learning unit 120 is not limited to that shown in formula (1). The learning unit 120 may use a loss function including only the term γ2V out of the terms γ2V and γ3Q, or may use a loss function including only the term γ3Q. The learning unit 120 may also use a loss function including a term other than the time integral loss term of the membrane potential or the loss term related to the firing condition as a term that indicates a better evaluation as the probability of each neuron model firing decreases.

[0031] (Time integral loss of membrane potential) The subformula V is expressed as formula (3).

[0032]

number

[0033] l indicates the identification number of the layer of the neural network. i indicates the identification number of the neuron model included in the l-th layer. player (l) is a constant for adjusting the degree of influence of the state of the lth layer, and player (l) >0. player (l) The value of may be preset by a user, for example. V (l) i is expressed as equation (4).

[0034]

number

[0035] V th represents the threshold membrane potential. The membrane potential of the neuron model is the threshold membrane potential V th When this is reached, the neuron model fires. v^ is 0 <v^<V th is a constant. The value of v^ corresponds to an example of a set value for the membrane potential. The time interval [0,T] is the time interval for firing of neuron models. Each neuron model fires at most once during the time interval [0,T]. T is a constant T>0.

[0036] v (l) i (t) represents the membrane potential of the i-th neuron model in the l-th layer at time t. θ represents a step function: θ(t)=1 when t≧0, and θ(t)=0 when t<0. If the i-th neuron model in the l-th layer does not fire by time T, V(l) i =0.

[0037] 2 is a diagram showing an example of time integral of membrane potential, in which the horizontal axis of the graph in Fig. 2 represents time, and the vertical axis represents membrane potential. Line L111 represents the membrane potential v of the ith neuron model in the lth layer. (l) i Represents (t).

[0038] Time t^ (l) i is the membrane potential v (l) i This represents the time when the value of (t) reaches the constant v^. (l) i If the value of (t) increases or decreases and reaches the constant v^ multiple times, then at time t^ (l) i is the membrane potential v (l) i represents the time when the value of (t) first reaches the constant v^. The area A111 is the membrane potential v indicated by the line L111. (l) i This is the region where the value of (t) is equal to or greater than the value of the constant v^. The integral ∫0 shown in equation (4) T Σ i (v (l) i (t)-v^)θ(v (l) i (t)-v^)θ(t (l) i -t)dt is the membrane potential v (l) i When the value of (t) is equal to or greater than the constant v^, the membrane potential v (l) i The value v obtained by subtracting the constant v^ from (t) (l) i Represents the integral of (t)-v^.

[0039] membrane potential v (l) i If the value of (t) is equal to or greater than the constant v^, the membrane potential v (l) i The value of (t) is the threshold membrane potential Vth This can be considered to be relatively close to the value of v, and it can be evaluated that the possibility of the neuron model firing is relatively high. The larger the value of the integral shown in equation (4), which is exemplified by the area of ​​region A111, the greater the membrane potential v (l) i The time that the value of (t) is greater than or equal to the constant v^ is long, or the membrane potential v (l) i The value of (t) is the threshold membrane potential V th Therefore, it can be evaluated that the larger the integral value shown in equation (4), the more likely the neuron model will fire. subformula V (l) i is the membrane potential v of the neuron model (l) i This is an example of a subexpression in which the smaller the value of the time integral with respect to (t) the better the evaluation.

[0040] Here, the value of the constant v^ is the membrane potential v (l) i The closer to the value of (t), the smaller the integral value shown in equation (4). Therefore, in equation (4), the value of the constant v^ is the membrane potential v (l) i To be able to properly evaluate the possibility of the neuron model firing even when the value of (t) is close to th -v^) value.

[0041] subformula V (l) i is the membrane potential v of a firing neuron model. (l) i (t) is greater than or equal to the constant v^ and the threshold potential V th At the time when the membrane potential v (l) i The integral of the difference between (t) and the constant v^ is the threshold membrane potential V th This is an example of a subexpression that shows the sum of the values ​​divided by the difference between and a constant v^. The value of the constant v^ is the threshold membrane potential V th This corresponds to an example of a setting value that is smaller than

[0042] In the calculation of equation (4), the membrane potential v (l) i If the membrane potential value is expressed as an analog value such as a current value or a voltage value, a mechanism for measuring the membrane potential value for each neuron model is required, which places a large burden on the hardware configuration. So, the membrane potential v (l) i Without needing to obtain the value of (t), we can (l) i We consider making it possible to calculate the value of

[0043] First, let us define the value of the constant v^ as the threshold voltage V th The limit v^→V th In particular, the value of the constant v^ is taken as the threshold membrane potential V th By making the value of (l) i The spikes input to the neuron model by (l) i The time t^ when (t) reaches the constant v^ (l) i is input to the neuron model before In this case, the loss function C is the constant v^, which is the threshold membrane potential V th This corresponds to an example of an evaluation function that takes the limit approaching

[0044] In this case, the time interval [t^ (l) i ,t (l) i During this time interval, no spikes are input to the neuron model, and the membrane potential v (l) i (t) may be monotonically increasing or decreasing. In particular, when this neuron model fires, v^ <V th So, the membrane potential v (l) i (t) increases monotonically. Thus, the membrane potential v (l) iWhen (t) increases monotonically, equation (4) can be transformed into equation (5).

[0045]

number

[0046] The subexpression v^(t (l) i -t^ (l) i ) is a constant and does not participate in learning, so this subexpression can be ignored. Therefore, equation (5) is replaced with equation (6).

[0047]

number

[0048] membrane potential v (l) i (t) is weighted by the weight coefficient w depending on the type of neuron model (type of spiking neuron). (l) ij And the firing time t of the neuron model in the previous layer (l-1) j This allows us to express the equation (6) as the membrane potential v (l) i This can be transformed into an equation in which (t) does not appear explicitly.

[0049] In the following, the time constant of the membrane potential is τ v and the time constant of the input current is τ I where τ is a constant > 0, and v =∞, and τ I If τ = ∞, v =∞, and τ I If τ = τ, then v =2τ I For each of the cases where =2τ, the subformula V (l) i Here is an example of the formula for τ v The value of may be preset by a user, for example. However, equation (6) is not limited to these neuron models and can be applied to various integrate-and-fire spiking neurons.

[0050] (τ v =∞, and τ I =∞) τ v =∞, and τ I If =∞, the membrane potential v of the ith neuron model in the lth layer (l) i is expressed as equation (7).

[0051]

number

[0052] j indicates the identification number of the neuron model included in the l-1th layer. N (l-1) represents the number of neuron models included in the l-1th layer. w (l) ij represents the weighting coefficient for the signal from the jth neuron model in the l-1th layer to the ith neuron model in the lth layer. Firing time t (l) i is expressed as equation (8).

[0053]

number

[0054] Gamma (l) i represents a set of identification numbers of neuron models that output a spike before the i-th neuron model in the l-1-th layer fires, among the neuron models included in the l-1-th layer. membrane potential v (l) i The time t^ reaches a constant v^ (l) i is expressed as equation (9).

[0055]

number

[0056] From equations (8) and (9), t (l) i -t^ (l) i is expressed as in equation (10).

[0057]

number

[0058] Equation (6) can be transformed into equation (11).

[0059]

number

[0060] Using equation (10), equation (11) can be transformed into equation (12).

[0061]

number

[0062] The value of the constant v^ is the threshold V th Close enough to t^ (l) i =t (l) i If we consider this, equation (12) can be transformed into equation (13).

number

[0063] In equation (13), the sub-expression V (l) i is the membrane potential v (l) iTherefore, according to the formula (13), the learning unit 120 can calculate the membrane potential v (l) i It is possible to calculate the time integral loss of the membrane potential without the need to obtain the value of (t). This makes it unnecessary to provide a mechanism for measuring the membrane potential value for each neuron model, even if the membrane potential value is expressed as an analog value such as a current value or a voltage value. In this respect, the neural network system 100 can reduce the burden on the hardware configuration.

[0064] Here, in equation (7), the membrane potential v (l) i (t) is the weighting coefficient w for the spike input to the ith neuron model in layer l. (l) ij and the firing time t of the neuron model that outputs a spike to the neuron model. (l-1) j The sub-formula V in formula (13) is expressed as (l) i is the membrane potential v of the neuron model (l) i (t) is the weighting factor w (l) ij and the firing time t of the neuron model that outputs a spike to the neuron model. (l-1) j This corresponds to an example of a subexpression obtained by expressing

[0065] In addition, when the learning unit 120 calculates the derivative of the loss function C, the subexpression “Σ j∈Γ(l)i w (l) ij ” and the firing time of the i-th neuron model in the l-th layer, “t (l) i " is treated as a constant. This means that the integral time interval [t^ (l) i ,t (l) i ] corresponds to fixing it.

[0066] In this case, the differentiation performed by the learning unit 120 corresponds to an example of differentiating the evaluation function by treating as constants the firing time of the neuron model being evaluated for the possibility of firing, which is included in the formula representing the evaluation function, and the sum of the weights for the spikes input to the neuron model being evaluated for the possibility of firing, which is included in the denominator of the fraction in the formula representing the evaluation function. Here, the loss function C corresponds to an example of the evaluation function. (l) i " corresponds to an example of the firing time of the neuron model that is being evaluated for the possibility of firing. j∈Γ(l)i w (l) ij " is an example of the total weighting value for spikes input to the neuron model that is being evaluated for the possibility of firing, which is included in the denominator of the fraction in the equation representing the evaluation function.

[0067] (τ v =∞, and τ I = τ) τ v =∞, and τ I If = τ, the membrane potential v of the ith neuron model in the lth layer (l) i is expressed as in equation (14).

[0068]

number

[0069] Firing time t (l) i is expressed as in equation (15).

[0070]

number

[0071] membrane potential v (l) i The time t^ reaches a constant v^ (l) iis expressed as in equation (16).

[0072]

number

[0073] From equations (15) and (16), t (l) i -t^ (l) i is expressed as in equation (17).

[0074]

number

[0075] The value of the constant v^ is the threshold membrane potential V th If the value of is sufficiently close to that of Eq. (17), then equation (17) can be transformed into equation (18) using an approximation by Taylor expansion of the logarithmic function.

[0076]

number

[0077] Equation (6) can be transformed into equation (19).

[0078]

number

[0079] The value of the constant v^ is the threshold membrane potential V th is close enough to the value of t (l) i -t^ (l) i When the value of is sufficiently small compared with the value of τ, equation (19) can be transformed into equation (20).

[0080]

number

[0081] Using equation (18), equation (20) can be transformed into equation (21).

[0082]

number

[0083] Equation (21) can be transformed into equation (22).

[0084]

number

[0085] In equation (22), the sub-formula V (l) i is the membrane potential v (l) i Therefore, according to the equation (22), the learning unit 120 can calculate the membrane potential v (l) i It is possible to calculate the time integral loss of the membrane potential without the need to obtain the value of (t). This makes it unnecessary to provide a mechanism for measuring the membrane potential value for each neuron model, even if the membrane potential value is expressed as an analog value such as a current value or a voltage value. In this respect, the neural network system 100 can reduce the burden on the hardware configuration.

[0086] Here, in equation (14), the membrane potential v (l) i (t) is the weighting coefficient w for the spike input to the ith neuron model in layer l. (l) ij and the firing time t of the neuron model that outputs a spike to the neuron model. (l-1) j The sub-formula V in formula (22) is expressed as (l) i is the membrane potential v of the neuron model (l) i (t) is the weighting factor w(l) ij and the firing time t of the neuron model that outputs a spike to the neuron model. (l-1) j This corresponds to an example of a subexpression obtained by expressing

[0087] In addition, when the learning unit 120 calculates the derivative of the loss function C, the subexpression “Σ j∈Γ(l)i w (l) ij ” and the firing time of the i-th neuron model in the l-th layer, “t (l) i " is treated as a constant. This means that the integral time interval [t^ (l) i ,t (l) i ] corresponds to fixing it.

[0088] In this case, the differentiation performed by the learning unit 120 corresponds to an example of differentiating the evaluation function by treating as constants the firing time of the neuron model being evaluated for the possibility of firing, which is included in the formula representing the evaluation function, and the sum of the weights for the spikes input to the neuron model being evaluated for the possibility of firing, which is included in the denominator of the fraction in the formula representing the evaluation function. Here, the loss function C corresponds to an example of the evaluation function. (l) i " corresponds to an example of the firing time of the neuron model that is being evaluated for the possibility of firing. j∈Γ(l)i w (l) ij " is an example of the total weighting value for spikes input to the neuron model that is being evaluated for the possibility of firing, which is included in the denominator of the fraction in the equation representing the evaluation function.

[0089] (τ v =2τ I = 2τ) τ v =2τ I When = 2τ, the membrane potential v of the ith neuron model in the lth layer (l)i is expressed as in equation (23).

[0090]

number

[0091] Ignition condition V th =v (l) i (t (l) i ) and the firing time t (l) i satisfies the condition shown in equation (24).

[0092]

number

[0093] Equation (24) can be transformed into equation (25).

[0094]

number

[0095] Equation (25) can be expressed as equation (26).

[0096]

number

[0097] a (l) i is expressed as in equation (27).

[0098]

number

[0099] b (l) i is expressed as in equation (28).

[0100]

number

[0101] Firing time t (l) i Using the formula for solving quadratic equations, equation (29) can be obtained from equation (25).

[0102]

number

[0103] Here, the other solution to the quadratic equation is the membrane potential v (l) i is the threshold V th From the larger value, the threshold V th It is ignored because it is the solution when it decreases to a smaller value. Equation (30) can be obtained from equation (29).

[0104]

number

[0105] Equation (30) can be transformed into equation (31).

number

[0106] In order to prevent deterioration of calculation accuracy due to cancellation of significant digits, it is possible to use equation (30) instead of equation (31) in implementing the neuron model.

[0107] membrane potential v (l) i The time t^ reaches the constant v^ (l) i is expressed as in equation (32).

[0108]

number

[0109] From equations (30) and (32), t (l) i -t^ (l) i is expressed as in equation (33).

[0110]

number

[0111] v^ is treated as a variable indicating the membrane potential and the threshold V th , and the threshold V th By performing a first-order Taylor expansion of v^ around , equation (33) can be transformed into equation (34).

[0112]

number

[0113] Equation (34) can be transformed into equation (35).

[0114]

number

[0115] The value of v^ is the threshold V th If it is sufficiently close to, equation (35) can be transformed into equation (36) using an approximation by Taylor expansion of a logarithmic function.

[0116]

number

[0117] α (l) i is expressed as in equation (37).

[0118]

number

[0119] Equation (6) can be transformed into equation (38).

[0120]

number

[0121] Using the Taylor expansion approximation of an exponential function, equation (38) can be transformed into equation (39).

[0122]

number

[0123] Equation (39) can be expressed as equation (40).

[0124]

number

[0125] In equation (40), the sub-expression V (l) i is the membrane potential v (l) i Therefore, according to equation (40), the learning unit 120 can calculate the membrane potential v (l) i It is possible to calculate the time integral loss of the membrane potential without the need to obtain the value of (t). This makes it unnecessary to provide a mechanism for measuring the membrane potential value for each neuron model, even if the membrane potential value is expressed as an analog value such as a current value or a voltage value. In this respect, the neural network system 100 can reduce the burden on the hardware configuration.

[0126] Here, in equation (23), the membrane potential v (l) i (t) is the weighting coefficient w for the spike input to the ith neuron model in layer l. (l) ijand the firing time t of the neuron model that outputs a spike to the neuron model. (l-1) j The sub-formula V in formula (40) is expressed as (l) i is the membrane potential v of the neuron model (l) i (t) is the weighting factor w (l) ij and the firing time t of the neuron model that outputs a spike to the neuron model. (l-1) j This corresponds to an example of a subexpression obtained by expressing

[0127] In addition, when the learning unit 120 calculates the derivative of the loss function C, the firing time “t (l) i ", said Alpha (l) i and are treated as constants. This means that the integral time interval [t^ (l) i ,t (l) i ] corresponds to fixing it.

[0128] In this case, the differentiation performed by the learning unit 120 is based on the firing time and α (l) i This is an example of differentiating the evaluation function by treating and as constants. Here, the loss function C is an example of an evaluation function. (l) i " is an example of a firing time for a neuron model that is being evaluated for its firing probability.

[0129] (Losses related to ignition conditions) Subformula Q in formula (1) is expressed as formula (41).

[0130]

number

[0131] If the loss function C includes both sub-formula V and sub-formula Q, then the player (l) and the value of player in the subexpression Q (l) The values ​​of may be set to the same value or may be set to different values. Q (l) i is expressed as in equation (42).

[0132]

number

[0133] t ref indicates the reference time for the firing time of the final layer (output layer). That is, at time t ref Any subsequent firings are ignored in the processing performed by spiking neural network 110.

[0134] 3 is a diagram showing an example of a firing condition of a neuron model, in which the horizontal axis of the graph in Fig. 3 represents time, and the vertical axis represents membrane potential. Line L211 represents the membrane potential v of the ith neuron model in the lth layer. (l) i Represents (t).

[0135] In the example in Figure 3, the firing time is time t (l) i The membrane potential indicated by the line L211 is the threshold membrane potential V th and the neuron model fires. Firing time t (l) i From now on, we assume that this neuron model does not accept input spikes and that it fires at time t (l) i After that, the membrane potential v (l) i Assuming that (t) changes, at the time of infinity, i.e., t→∞, the membrane potential becomes τ v Σ j∈Γ(l)i w (l)ij converges to. From this, equation (43) can be considered as the ignition condition.

[0136]

number

[0137] Σ j∈Γ(l)i w (l) ij It is believed that the smaller the value of Σ, the more difficult it is for the ignition condition shown in equation (43) to be satisfied, and ignition is suppressed. j∈Γ(l)i w (l) ij can be used as an index of the possibility that the neuron model fires. Using this index, we can define the loss (subformula Q) for the firing condition, which indicates a better evaluation as the neuron model is less likely to fire, as shown in the above equations (41) and (42).

[0138] The sub-formula Q shown in equations (41) and (42) is an example of a sub-formula that, for a fired neuron model, the smaller the total value of the weighting coefficients for the spikes input to the neuron model, the better the evaluation. The timing at which the neuron model is enabled to fire is within a finite time interval [0,t ref ], so t (l) i <t ref The conditions are set as follows.

[0139] (About the experimental results) An experiment was conducted to confirm the operation of the neural network system 100. First, we trained a spiking neural network using a loss function that includes a term γ2V for the time integral loss of the membrane potential, and observed the firing status of the neuron model. Here, we used a three-layer spiking neural network with an input layer containing 784 neuron models, a hidden layer containing 400 neuron models, and an output layer containing 10 neuron models. As for the hyperparameters, τv = τ I =∞, t ref = 8 (seconds), and the coefficient γ of the temporal penalty term is γ = 10 -4 The learning rate η is η = 10 -4 It was decided. Coefficient γ2 value 0, 2.3×10 -7 , 4.8×10 -6 For each of these, we trained character recognition using MNIST as training data.

[0140] Figure 4 shows an example of the firing state of the neuron model observed when γ2 = 0. Therefore, Figure 4 shows the firing state of the neuron model when the influence degree of the term γ2V of the time integral loss of the membrane potential is set to 0. The horizontal axis of the graph in Figure 4 represents time. The vertical axis represents the identification number of the neuron model, and the identification number of each neuron is assigned to the horizontal axis for each of the input layer, hidden layer, and output layer. The firing times of the neuron models are shown as dots (black circles). In addition, in the example of Figure 4, the recognition rate of the spiking neural network was 98.05%.

[0141] Figure 5 shows γ2 = 2.3 × 10 -7 FIG. 13 is a diagram showing an example of the firing state of a neuron model observed in the case of The horizontal axis of the graph in Figure 5 represents time. The vertical axis represents the identification number of the neuron model, and the identification number of each neuron is assigned to the horizontal axis for each of the input layer, hidden layer, and output layer. The firing times of the neuron models are shown as dots (black circles). In addition, in the example of Figure 5, the recognition rate of the spiking neural network was 97.84%.

[0142] In the example of Figure 5, the number of firings of the neuron model in the hidden layer is reduced compared to the example of Figure 4. Also, in terms of the recognition rate, the example of Figure 5 achieves a recognition rate that is comparable to that of Figure 4. In this way, in the example of FIG. 5, it is possible to reduce power consumption more than in the example of FIG. 4, and at the same time, a recognition rate as in the example of FIG. 4 is obtained.

[0143] Figure 6 shows γ2 = 4.8 × 10 -6 6 shows an example of the firing state of the neuron model observed in the case of . Therefore, Fig. 6 shows the firing state of the neuron model when the influence degree of the term γ2V of the time integral loss of the membrane potential is further strengthened (increased) than in the case of Fig. 5. The horizontal axis of the graph in Figure 6 represents time, and the vertical axis represents the identification number of the neuron model, with the identification numbers of each neuron assigned to the horizontal axis for the input layer, hidden layer, and output layer. The firing times of the neuron models are shown as dots (black circles). In addition, in the example of Figure 6, the recognition rate of the spiking neural network was 97.6%.

[0144] In the example of Figure 6, the number of firings of the neuron model in the hidden layer is further reduced compared to the example of Figure 5. Also, the recognition rate of the example of Figure 6 is about the same as that of the examples of Figures 4 and 5. In this way, in the example of FIG. 6, it is possible to reduce power consumption even further than in the example of FIG. 5, and at the same time, a recognition rate as in the example of FIG. 5 is obtained.

[0145] In this way, it was confirmed that the number of firings of the neuron model can be reduced by introducing the term γ2V, which is the time integral loss of the membrane potential. By reducing the number of firings of the neuron model, it is expected that the power consumption of the spiking neural network can be reduced. In addition, the recognition rate of the spiking neural network was the same even when the number of firings of the neuron model was reduced as when the number of firings of the neuron model was not reduced.

[0146] In addition, for the case where the term γ2V representing the time integral loss of the membrane potential was used, and for the case where the term γ3Q representing the loss related to the firing condition was used, the coefficients γ2 and γ3 were set to various values ​​to train the spiking neural network, and the relationship between the recognition rate of the neural network and the firing rate of the neuron model in the hidden layer was observed. The recognition rate here is calculated by dividing the number of correctly recognized data by the total number of data.

[0147] Here, a three-layer spiking neural network and a five-layer spiking neural network were used, respectively. The values ​​of the hyperparameters were the same as those in the experiment described with reference to Figs. 4 to 6. For the three-layer spiking neural network, a spiking neural network was used in which the input layer contained 784 neuron models, the hidden layer contained 400 neuron models, and the output layer contained 10 neuron models. For the five-layer spiking neural network, we used a spiking neural network in which the input layer contained 784 neuron models, the first hidden layer, the second hidden layer, and the third hidden layer each contained 400 neuron models, and the output layer contained 10 neuron models.

[0148] For training data, we used MNIST and Fashion-MNIST to learn character recognition. The firing rate of each neuron model in the repetition of the experiment was calculated as the firing rate of each neuron model. Specifically, the firing rate of each neuron model was calculated by dividing the number of firings by the number of repetitions of the experiment, and the average firing rate of each neuron model was calculated for all neuron models included in the hidden layer. In a configuration with three hidden layers, the average value was calculated for each hidden layer.

[0149] Fig. 7 is a diagram showing a first example of the relationship between the recognition rate of a neural network and the firing rate of a neuron model in a hidden layer. Fig. 7 shows an example in which a three-layer spiking neural network is used and MNIST is used as training data. The horizontal axis of the graph in Fig. 7 represents the firing rate, and the vertical axis represents the recognition rate.

[0150] Line L311 shows an example of the relationship between the recognition rate of the neural network and the firing rate of the neuron model in the hidden layer when the term γ2V, which is the time integral loss of the membrane potential, is used. Line L312 shows an example of the relationship between the recognition rate of the neural network and the firing rate of the neuron model in the hidden layer when the term γ3Q, which is the loss related to the firing condition, is used.

[0151] In the example of Figure 7, when the term γ2V, which is the time integral loss of the membrane potential, is used, and when the term γ3Q, which is the loss related to the firing condition, is used, the recognition rate is within the range of 97.6% to 98.0% for the firing rate range of 0.1 to 0.8. In this way, the firing rate can be reduced while roughly maintaining the recognition accuracy of the neural network.

[0152] Fig. 8 is a diagram showing a second example of the relationship between the recognition rate of a neural network and the firing rate of a neuron model in the intermediate layer. Fig. 8 shows an example in which a five-layer spiking neural network is used and MNIST is used as training data. The horizontal axis of the graph in Figure 8 indicates the firing rate, and the vertical axis indicates the recognition rate. Figure 8 shows the relationship between the firing rate and the recognition rate for each of the first hidden layer, the second hidden layer, and the third hidden layer, and for the average firing rate of the neuron models in these three hidden layers.

[0153] Line L411 shows an example of the relationship between the recognition rate of the neural network and the firing rate of the neuron model in the first hidden layer when the term γ2V, which is the time integral loss of the membrane potential, is used. Line L412 shows an example of the relationship between the recognition rate of the neural network and the firing rate of the neuron model in the first hidden layer when the term γ3Q, which is the loss related to the firing condition, is used.

[0154] Line L421 shows an example of the relationship between the recognition rate of the neural network and the firing rate of the neuron model in the second hidden layer when the term γ2V, which is the time integral loss of the membrane potential, is used. Line L422 shows an example of the relationship between the recognition rate of the neural network and the firing rate of the neuron model in the second hidden layer when the term γ3Q, which is the loss related to the firing condition, is used.

[0155] Line L431 shows an example of the relationship between the recognition rate of the neural network and the firing rate of the neuron model in the third hidden layer when the term γ2V, which is the time integral loss of the membrane potential, is used. Line L432 shows an example of the relationship between the recognition rate of the neural network and the firing rate of the neuron model in the third hidden layer when the term γ3Q, which is the loss related to the firing condition, is used.

[0156] Line L441 shows an example of the relationship between the recognition rate of the neural network and the average firing rate of the neuron models from the first hidden layer to the third hidden layer when the term γ2V, which is the time integral loss of the membrane potential, is used. Line L442 shows an example of the relationship between the recognition rate of the neural network and the average firing rate of the neuron models from the first hidden layer to the third hidden layer when the term γ3Q, which is the loss related to the firing condition, is used.

[0157] In the example of Figure 8, when the term γ2V for the time integral loss of the membrane potential is used, when the term γ3Q for the loss related to the firing condition is used, and when the firing rate is 0.2 or more, for the first hidden layer, the second hidden layer, the third hidden layer, and the average value, a recognition rate in the range of about 97.5% to about 98.0% is obtained. In this way, the firing rate can be reduced while roughly maintaining the recognition accuracy of the neural network.

[0158] Fig. 9 is a diagram showing a third example of the relationship between the recognition rate of a neural network and the firing rate of a neuron model in the intermediate layer. Fig. 9 shows an example in which a three-layer spiking neural network is used and Fashion-MNIST is used as training data. The horizontal axis of the graph in Fig. 9 represents the firing rate, and the vertical axis represents the recognition rate.

[0159] Line L511 shows an example of the relationship between the recognition rate of the neural network and the firing rate of the neuron model in the hidden layer when the term γ2V, which is the time integral loss of the membrane potential, is used. Line L512 shows an example of the relationship between the recognition rate of the neural network and the firing rate of the neuron model in the hidden layer when the term γ3Q, which is the loss related to the firing condition, is used.

[0160] In the example of Figure 9, when the term γ2V, which is the time integral loss of the membrane potential, is used, and when the term γ3Q, which is the loss related to the firing condition, is used, the recognition rate is within the range of 87.00% to 88.00% in the section of the firing rate from 0.1 to 0.3. In this way, the firing rate was reduced while roughly maintaining the recognition accuracy of the neural network.

[0161] Fig. 10 is a diagram showing a fourth example of the relationship between the recognition rate of a neural network and the firing rate of a neuron model in a hidden layer. Fig. 10 shows an example in which a five-layer spiking neural network is used and Fashion-MNIST is used as training data. The horizontal axis of the graph in Fig. 10 indicates the firing rate, and the vertical axis indicates the recognition rate. Fig. 10 shows the relationship between the firing rate and the recognition rate for each of the first hidden layer, the second hidden layer, and the third hidden layer, and for the average firing rate of the neuron models in these three hidden layers.

[0162] Line L611 shows an example of the relationship between the recognition rate of the neural network and the firing rate of the neuron model in the first hidden layer when the term γ2V, which is the time integral loss of the membrane potential, is used. Line L612 shows an example of the relationship between the recognition rate of the neural network and the firing rate of the neuron model in the first hidden layer when the term γ3Q, which is the loss related to the firing condition, is used.

[0163] Line L621 shows an example of the relationship between the recognition rate of the neural network and the firing rate of the neuron model in the second hidden layer when the term γ2V, which is the time integral loss of the membrane potential, is used. Line L622 shows an example of the relationship between the recognition rate of the neural network and the firing rate of the neuron model in the second hidden layer when the term γ3Q, which is the loss related to the firing condition, is used.

[0164] Line L631 shows an example of the relationship between the recognition rate of the neural network and the firing rate of the neuron model in the third hidden layer when the term γ2V, which is the time integral loss of the membrane potential, is used. Line L632 shows an example of the relationship between the recognition rate of the neural network and the firing rate of the neuron model in the third hidden layer when the term γ3Q, which is the loss related to the firing condition, is used.

[0165] Line L641 shows an example of the relationship between the recognition rate of the neural network and the average firing rate of the neuron models from the first hidden layer to the third hidden layer when the term γ2V, which is the time integral loss of the membrane potential, is used. Line L642 shows an example of the relationship between the recognition rate of the neural network and the average firing rate of the neuron models from the first hidden layer to the third hidden layer when the term γ3Q, which is the loss related to the firing condition, is used.

[0166] In the example of Figure 10, when the term γ2V for the time integral loss of the membrane potential is used, when the term γ3Q for the loss related to the firing condition is used, and when the firing rate is 0.2 or more for the first hidden layer, the second hidden layer, the third hidden layer, and the average value, a recognition rate in the range of about 87% to about 89% is obtained. In this way, the firing rate can be reduced while roughly maintaining the recognition accuracy of the neural network.

[0167] As described above, the spiking neural network 110 is a time-based spiking neural network configured using neuron models based on an integrate-and-fire model. The learning unit 120 performs learning of the spiking neural network using an evaluation function that indicates a better evaluation the lower the probability that each neuron model fires. As a result, in the neural network system 100, the spiking neural network 110 can be trained to suppress firing of the neuron model, and in this respect, it is expected that the power consumption of the spiking neural network 110 during operation will be relatively small.

[0168] Moreover, the learning unit 120 performs learning of the spiking neural network using a loss function C including a sub-formula V indicating a time integral loss of the membrane potential. The loss function C including a sub-formula V indicating a time integral loss of the membrane potential corresponds to an example of an evaluation function that indicates a better evaluation as the value of the time integral of the membrane potential of the neuron model is smaller. As a result, in the neural network system 100, the spiking neural network 110 can be trained to suppress firing of the neuron model, and in this respect, it is expected that the power consumption of the spiking neural network 110 during operation will be relatively small.

[0169] In addition, the learning unit 120 calculates the membrane potential v (l) i (t) is the threshold voltage V th The constant v^ is smaller than the threshold membrane potential V th At the time when the membrane potential v (l) i The integral of the difference between (t) and the constant v^ is the threshold membrane potential V th The spiking neural network is trained using a loss function C that contains a subexpression V that indicates the sum of values ​​obtained by dividing by the difference between and a constant v^.

[0170] Here, the membrane potential v(l) i If the value of (t) is equal to or greater than the constant v^, the membrane potential v (l) i The value of (t) is the threshold membrane potential V th This means that the neuron model is likely to fire. (l) i (t) is the threshold voltage V th The constant v^ is smaller than the threshold membrane potential V th At the time when the membrane potential v (l) i The larger the integral of the difference between (t) and the constant v^, the greater the membrane potential v (l) i The time that the value of (t) is greater than or equal to the constant v^ is long, or the membrane potential v (l) i The value of (t) is the threshold membrane potential V th The learning unit 120 can use the value of this integral to evaluate the likelihood that the neuron model will fire.

[0171] Also, the value of the constant v^ is the membrane potential v (l) i The closer the value of the integral is to the value of (t), the smaller the value of the integral becomes. The learning unit 120 calculates the value of the integral as the threshold membrane potential V th By dividing the value by the difference between v^ and the constant v^, the value of the constant v^ is calculated as the membrane potential v (l) i It is expected that the probability of the neuron model firing can be appropriately evaluated even when the value is close to (t).

[0172] In addition, the learning unit 120 calculates the constant v^ based on the threshold membrane potential V th The spiking neural network is trained by a learning method that uses an evaluation function that takes the limit approaching This allows the subformula V to be written as in the examples of formulas (5) and (6) above. (l) i This simplifies the above, and reduces the calculation load imposed on the learning unit 120 when differentiating the loss function C.

[0173] In addition, the learning unit 120 calculates the membrane potential v (l) i (t) is the weighting coefficient Σ for the spikes input to the neuron model. j∈Γ(l)i w (l) ij and the firing time t of the neuron model that outputs a spike to the neuron model. (l-1) j The spiking neural network 110 is trained using a loss function C obtained by expressing

[0174] As a result, the sub-formula V, as in the example of formula (13), the example of formula (23), and the example of formula (40), (l) i can be a sub-expression where the membrane potential does not appear explicitly, and the loss function C can be a loss function where the membrane potential does not appear explicitly. According to the neural network system 100, even if the value of the membrane potential is expressed as an analog value such as a current value or a voltage value, there is no need to set up a mechanism for measuring the value of the membrane potential for each neuron model, which reduces the burden on the hardware configuration.

[0175] In addition, the learning unit 120 detects the membrane potential v (l) i The time t^ when (t) reaches the constant v^ (l) i Then, the firing time t (l) i Time interval until [t^ (l) i ,t (l) i ] is treated as a fixed time interval and the loss function C is differentiated. This allows the differentiation calculations performed by the learning unit 120 to be simplified, and in this respect, the calculation load on the learning unit 120 can be reduced.

[0176] In addition, the learning unit 120 trains the spiking neural network 110 using a loss function C including a subexpression Q that indicates a better evaluation for a fired neuron model when the total value of the weighting coefficients for the spikes input to the neuron model is smaller. As a result, in the neural network system 100, the spiking neural network 110 can be trained to suppress firing of the neuron model, and in this respect, it is expected that the power consumption of the spiking neural network 110 during operation will be relatively small.

[0177] In addition, the learning unit 120 learns a sub-formula Σ that shows a better evaluation as the estimation accuracy of the spiking neural network 110 is higher. p κ p lnS p The spiking neural network 110 is trained using a loss function C including sub-equations V and Q that show a better evaluation as the probability of each neuron model firing decreases, and coefficients γ2 and γ3 for adjusting the degree of influence of the sub-equations V and Q that show a better evaluation as the probability of each neuron model firing decreases. As a result, in neural network system 100, the trade-off between the estimation accuracy of spiking neural network 110 and the power consumption of spiking neural network 110 can be adjusted by the simple process of adjusting the values ​​of coefficients γ2 and γ3.

[0178] 11 is a diagram showing an example of the configuration of a neural network system according to an embodiment. In the configuration shown in FIG. 11, a neural network system 610 includes a spiking neural network 611 and a learning unit 612. In this configuration, the spiking neural network 611 is a time-based spiking neural network configured using neuron models based on an integrate-and-fire model. The learning unit 612 learns the spiking neural network using an evaluation function that indicates a better evaluation the lower the probability that each neuron model fires. The learning unit 612 corresponds to an example of a learning means.

[0179] As a result, in the neural network system 610, the spiking neural network 611 can be trained to suppress firing of the neuron model, and in this respect, it is expected that the power consumption of the spiking neural network 611 during operation will be relatively small.

[0180] 12 is a diagram showing an example of the configuration of a learning device according to an embodiment. In the configuration shown in FIG. In this configuration, the learning unit 621 learns the spiking neural network using an evaluation function that indicates a better evaluation the lower the possibility that each neuron model in the time-based spiking neural network, which is configured using neuron models based on an integrate-and-fire model, fires. The learning unit 621 corresponds to an example of a learning means.

[0181] According to the learning device 620, it is possible to train a spiking neural network so as to suppress firing of a neuron model, and in this respect, it is expected that the power consumption during operation of the spiking neural network will be relatively small.

[0182] Fig. 13 is a diagram showing an example of a processing procedure in a learning method according to the embodiment. The learning method shown in Fig. 13 includes performing learning (step S611). In the learning step (step S611), the computer learns the spiking neural network using an evaluation function that indicates a better evaluation the lower the probability that each neuron model in the time-based spiking neural network, which is configured using neuron models based on an integrate-and-fire model, fires.

[0183] According to the learning method shown in FIG. 13, a spiking neural network can be trained so as to suppress firing of the neuron model. In this respect, it is expected that the power consumption during operation of the spiking neural network will be relatively small.

[0184] FIG. 14 illustrates an example configuration of a computer in accordance with at least one embodiment. In the configuration shown in FIG. 14, a computer 700 includes a CPU 710, a main memory device 720, an auxiliary memory device 730, an interface 740, and a non-volatile recording medium 750.

[0185] Any one or more of the neural network system 100, the neural network system 610, and the learning device 620, or a part of them, may be implemented in the computer 700. In this case, the operation of each of the above-mentioned processing units is stored in the auxiliary storage device 730 in the form of a program. The CPU 710 reads the program from the auxiliary storage device 730, develops it in the main storage device 720, and executes the above-mentioned processing according to the program. The CPU 710 also secures a storage area corresponding to each of the above-mentioned storage units in the main storage device 720 according to the program. Communication between each device and other devices is executed by the interface 740 having a communication function and performing communication according to the control of the CPU 710. The interface 740 also has a port for the nonvolatile recording medium 750, and reads information from the nonvolatile recording medium 750 and writes information to the nonvolatile recording medium 750.

[0186] When the neural network system 100 is implemented in a computer 700, the operations of the spiking neural network 110 and the learning unit 120 are stored in the form of a program in the auxiliary storage device 730. The CPU 710 reads the program from the auxiliary storage device 730, loads it into the main storage device 720, and executes the above-mentioned processing according to the program.

[0187] Furthermore, the CPU 710 allocates a storage area in the main storage device 720 for the neural network system 100 to perform processing in accordance with the program. Communication between the neural network system 100 and other devices is performed by the interface 740 having a communication function and operating under the control of the CPU 710. Interaction between the neural network system 100 and a user is performed by the interface 740 having an input device and an output device, presenting information to the user via the output device under the control of the CPU 710, and accepting user operations via the input device.

[0188] In addition, the spiking neural network 110 may be configured using an application specific integrated circuit (ASIC), a field programmable gate array (FGPA), or a circuit using analog devices, or a combination of these. In this case, the spiking neural network 110 may be configured as a part of the interface 740 and may operate under the control of the CPU 710. Alternatively, spiking neural network 110 may be configured as a device or system separate from computer 700 and operate according to a program for spiking neural network 110. In this case, spiking neural network 110 may be configured to communicate with computer 700. Alternatively, both the spiking neural network 110 and the learning unit 120 may be configured using hardware including an ASIC or an FGPA, and may operate according to a program for the spiking neural network 110.

[0189] When the neural network system 610 is implemented in the computer 700, the operations of the spiking neural network 611 and the learning unit 612 are stored in the form of a program in the auxiliary storage device 730. The CPU 710 reads the program from the auxiliary storage device 730, loads it in the main storage device 720, and executes the above-mentioned processing according to the program.

[0190] Furthermore, the CPU 710 allocates a storage area in the main storage device 720 for the neural network system 610 to perform processing in accordance with the program. Communication between the neural network system 610 and other devices is performed by the interface 740 having a communication function and operating under the control of the CPU 710. Interaction between the neural network system 610 and a user is performed by the interface 740 having an input device and an output device, presenting information to the user via the output device under the control of the CPU 710, and accepting user operations via the input device.

[0191] The spiking neural network 611 may be configured using an application specific integrated circuit (ASIC), a field programmable gate array (FGPA), or a circuit using analog devices, or a combination of these. In this case, the spiking neural network 611 may be configured as a part of the interface 740 and may operate under the control of the CPU 710. Alternatively, the spiking neural network 611 may be configured as a device or system separate from the computer 700 and may operate according to a program for the spiking neural network 611. In this case, the spiking neural network 611 may be configured to communicate with the computer 700. Alternatively, both the spiking neural network 611 and the learning unit 612 may be configured using hardware including an ASIC or an FGPA, and may operate according to a program for the spiking neural network 110.

[0192] When the learning device 620 is implemented in the computer 700, the operation of the learning unit 621 is stored in the form of a program in the auxiliary storage device 730. The CPU 710 reads the program from the auxiliary storage device 730, loads it in the main storage device 720, and executes the above-mentioned processing according to the program.

[0193] Furthermore, the CPU 710 allocates a storage area in the main storage device 720 for the learning device 620 to perform processing in accordance with the program. Communication between learning device 620 and other devices is performed by interface 740 having a communication function and operating under the control of CPU 710. Interaction between learning device 620 and a user is performed by interface 740 having an input device and an output device, presenting information to the user via the output device under the control of CPU 710, and accepting user operations via the input device.

[0194] Any one or more of the above-mentioned programs may be recorded in non-volatile recording medium 750. In this case, interface 740 may read the program from non-volatile recording medium 750. Then, CPU 710 may directly execute the program read by interface 740, or may temporarily store the program in main storage device 720 or auxiliary storage device 730 and execute it.

[0195] Note that the programs for executing all or part of the processing performed by neural network system 100, neural network system 610, and learning device 620 may be recorded on a computer-readable recording medium, and the programs recorded on the recording medium may be read into a computer system and executed to perform the processing of each part. Note that the term "computer system" here includes hardware such as an OS and peripheral devices. Furthermore, the term "computer-readable recording medium" refers to portable media such as flexible disks, optical magnetic disks, ROMs (Read Only Memory), and CD-ROMs (Compact Disc Read Only Memory), as well as storage devices such as hard disks built into computer systems. The above-mentioned program may be for realizing part of the above-mentioned functions, or may be capable of realizing the above-mentioned functions in combination with a program already recorded in the computer system.

[0196] Although an embodiment of the present invention has been described in detail above with reference to the drawings, the specific configuration is not limited to this embodiment, and designs that do not deviate from the gist of the present invention are also included.

[0197] A part or all of the above-described embodiments can be described as, but is not limited to, the following supplementary notes.

[0198] (Appendix 1) A time-based spiking neural network constructed using an integrate-and-fire neuron model; A learning means for learning the spiking neural network using an evaluation function that indicates a better evaluation as the probability of each neuron model firing decreases; A neural network system comprising:

[0199] (Appendix 2) the learning means performs learning of the spiking neural network using the evaluation function which indicates a better evaluation as the value of the time integral of the membrane potential of the neuron model is smaller; 2. The neural network system of claim 1.

[0200] (Appendix 3) the learning means performs learning of the spiking neural network using the evaluation function including a sub-expression indicating a sum of values ​​obtained by dividing an integral of a difference between a membrane potential of a neuron model that has fired and a set value, the integral being a value smaller than a threshold membrane potential, and a value equal to or larger than the threshold membrane potential, by a difference between the threshold membrane potential and the set value, during a time period when the membrane potential of the neuron model is equal to or larger than the set value, the value being smaller than the threshold membrane potential, and is equal to or smaller than the threshold membrane potential; 3. The neural network system of claim 2.

[0201] (Appendix 4) the learning means performs learning of the spiking neural network by a learning method of taking a differential of the evaluation function using the evaluation function which takes a limit that makes the set value approach the threshold membrane potential; 4. The neural network system of claim 3.

[0202] (Appendix 5) the learning means performs learning of the spiking neural network using the evaluation function obtained by expressing the membrane potential of the neuron model using a weighting coefficient for a spike input to the neuron model and a firing time of a neuron model which outputs a spike to the neuron model; 5. The neural network system of claim 4.

[0203] (Appendix 6) the learning means differentiates the evaluation function by treating a time interval from the time when the membrane potential reaches the set value to the time when the neuron model, which is the subject of evaluation of the possibility of firing, as a fixed time interval. 6. The neural network system of claim 4 or 5.

[0204] (Appendix 7) the learning means performs learning of the spiking neural network using the evaluation function which indicates a better evaluation for a fired neuron model as the total value of weighting coefficients for spikes input to the neuron model is smaller; 7. A neural network system according to any one of claims 1 to 6.

[0205] (Appendix 8) the learning means performs learning of the spiking neural network using the evaluation function including a sub-equation showing a better evaluation as the estimation accuracy of the spiking neural network is higher, a sub-equation showing a better evaluation as the possibility of each neuron model firing is lower, and a coefficient for adjusting the degree of influence of the sub-equation showing a better evaluation as the possibility of each neuron model firing is lower; 8. The neural network system of claim 1.

[0206] (Appendix 9) a learning means for learning a time-based spiking neural network configured using neuron models based on an integrate-and-fire model, using an evaluation function that indicates a better evaluation as the probability of each neuron model firing decreases; A learning device comprising:

[0207] (Appendix 10) The computer A spiking neural network of a time system configured using neuron models based on an integrate-and-fire model is trained using an evaluation function that indicates a better evaluation the lower the possibility of each neuron model firing. A learning method that includes:

[0208] (Appendix 11) On the computer, A spiking neural network of a time-based system configured using neuron models based on an integrate-and-fire model is trained using an evaluation function that indicates a better evaluation the lower the possibility of each neuron model firing. A program for executing the above. [Explanation of symbols]

[0209] 100, 610 Neural network system 110, 611 Spiking Neural Networks 120, 612, 621 Learning Department 620 Learning Device

Claims

1. A time-based spiking neural network constructed using an integrate-and-fire neuron model; A learning means for learning the spiking neural network using an evaluation function that indicates a better evaluation as the probability of each neuron model firing decreases; A neural network system comprising:

2. the learning means performs learning of the spiking neural network using the evaluation function which indicates a better evaluation as the value of the time integral of the membrane potential of the neuron model is smaller; 2. The neural network system according to claim 1.

3. the learning means performs learning of the spiking neural network using the evaluation function including a sub-expression indicating a sum of values ​​obtained by dividing an integral of a difference between a membrane potential of a neuron model that has fired and a set value, the integral being a value smaller than a threshold membrane potential, and a value equal to or larger than the threshold membrane potential, by a difference between the threshold membrane potential and the set value, during a time period when the membrane potential of the neuron model is equal to or larger than the set value, the value being smaller than the threshold membrane potential, and is equal to or smaller than the threshold membrane potential; 3. The neural network system according to claim 2.

4. the learning means performs learning of the spiking neural network by a learning method of taking a differential of the evaluation function using the evaluation function which takes a limit that makes the set value approach the threshold membrane potential; 4. The neural network system according to claim 3.

5. the learning means performs learning of the spiking neural network using the evaluation function obtained by expressing the membrane potential of the neuron model using a weighting coefficient for a spike input to the neuron model and a firing time of a neuron model which outputs a spike to the neuron model; 5. The neural network system according to claim 4.

6. the learning means differentiates the evaluation function by treating a time interval from the time when the membrane potential reaches the set value to the time when the neuron model fires as a fixed time interval; 6. The neural network system according to claim 4 or 5.

7. the learning means performs learning of the spiking neural network using the evaluation function which indicates a better evaluation for a fired neuron model as the total value of weighting coefficients for spikes input to the neuron model is smaller; 2. The neural network system according to claim 1.

8. a learning means for learning a time-based spiking neural network configured using neuron models based on an integrate-and-fire model, using an evaluation function that indicates a better evaluation as the probability of each neuron model firing decreases; A learning device comprising:

9. The computer A spiking neural network of a time system configured using neuron models based on an integrate-and-fire model is trained using an evaluation function that indicates a better evaluation the lower the possibility of each neuron model firing. A learning method that includes:

10. On the computer, A spiking neural network of a time-based system configured using neuron models based on an integrate-and-fire model is trained using an evaluation function that indicates a better evaluation the lower the possibility of each neuron model firing. A program for executing.

Citation Information

Patent Citations

  • Constructing a neural network for a low spiking rate

    JP2017509953A